354results about How to "Avoid vanishing gradients" patented technology

The invention discloses a fundus image retinal vessel segmentation method and a fundus image retinal vessel segmentation system based on deep learning. The fundus image retinal vessel segmentation method comprises the steps of performing data amplification on a training set, enhancing an image, training a convolutional neural network by using the training set, segmenting the image by using a convolutional neural network segmentation model to obtain a segmentation result, training a random forest classifier by using features of the convolutional neural network, extracting a last layer of convolutional layer output from the convolutional neural network, using the convolutional layer output as input of the random forest classifier for pixel classification to obtain another segmentation result, and fusing the two segmentation results to obtain a final segmentation image. Compared with the traditional vessel segmentation method, the fundus image retinal vessel segmentation method uses the deep convolutional neural network for feature extraction, the extracted features are more sufficient, and the segmentation precision and efficiency are higher.

The invention discloses a generative adversarial network-based grayscale picture colorizing method. A DiscoGAN, a Progressive Growing GAN, a Wasserstein GAN and a CGAN are combined to generate a generative adversarial network. The method comprises the following steps of: firstly collecting and arranging picture samples and dividing the samples into two groups, wherein one group comprises N grayscale pictures and the other group comprises N color pictures; designing architecture of the generative adversarial network to ensure that the trained network can generate pictures with high resolutionsand high quality; transmitting the samples into the generative adversarial network to start training, and after the generative adversarial network is stably trained, enhancing the resolutions of generated pictures by using a PGGAN. According to the method, WGAN-PG is added in the network to improve the original generative adversarial network, so that the problems of gradient instability and mode collapse are solved, and the process of optimizing the generative adversarial network is improved. Finally, a description limiting function of the CGAN is added in the network, so that pictures with appointed styles can be generated according to descriptions.

The present invention discloses a series-wound long short-term memory recurrent neural network-based heating load prediction method. The method comprises the steps of constructing a sample data set based on temperature, climate and heat supply data during a given period of time, and respectively subjecting the input data and the output data of the sample data set to standardized treatment; dividing the input data into two portions, respectively inputting the two portions into two independent long short-term memory recurrent neural networks to merge the two portions of the input data, inputting the output data to a long short-term memory recurrent neural network at a next layer, and finally inputting the data into two full connection layers; training a constructed series-wound long short-term memory recurrent neural network, and optimizing the network by adopting the parameter optimization-based adaptive torque estimation algorithm; inputting to-be-predicted data into the series-wound long short-term memory recurrent neural network, calculating and obtaining a heating load prediction result. The method of the invention can effectively discriminate input data, and accelerate the learning speed. Therefore, the learning efficiency is improved and the prediction accuracy is improved.

The invention discloses an aircraft track prediction method based on a depth memory network. Firstly, the method extracts the historical aircraft track data and preprocesses the data to form an aircraft track information data sequence under a time sequence. On this basis, the input and output sample characteristics are constructed, and the samples are standardized to generate dimensionless training data set. Then, the attention mechanism model is integrated into the long-term and short-term memory network structure, and the structural parameters are initialized to construct the track prediction neural network. Finally, the track prediction network is trained and optimized, and the track prediction network model with high prediction accuracy is obtained.

The invention provides a medical image segmentation method based on an improved convolutional neural network, and belongs to the field of image processing. According to the method, an improved convolutional neural network is constructed, and a multi-channel fusion module, a multi-channel dense connection module, a down-sampling module and an up-sampling module are designed on the basis of an original framework of a standard U-Net network. The depth of the network is increased while the network training calculation amount is controlled, and redundant calculation is reduced. Gradient disappearance is avoided by improving the internal structure of the network while the network is deepened. The experimental result is that the segmentation DICE coefficient of the output prediction graph of thetrained neural network is 98.57%, the DICE coefficient segmented by the original network is 98.26%, and the effectiveness of the method is reflected.

The invention discloses a mobile pollution source emission concentration prediction method based on deep learning, and provides a convolutional long-short-term memory neural network prediction methodbased on an attention mechanism according to regional space-time distribution characteristics of mobile pollution source pollutants. Firstly, a Granger causal relationship between stations is analyzedand a hyper-parameter Gaussian vector weight function is developed to determine a spatial autocorrelation variable as a part of an input feature; secondly, extracting time-space characteristics of data used by the LSTM network by using a convolutional neural network, and meanwhile, attention models are respectively used for weighting a characteristic graph and a channel so as to enhance the effectiveness of the characteristics; finally, a time series predictor based on deep LSTM is used to learn long-term and short-term dependency of the atmospheric pollutant concentration. According to the method, inherent useful characteristics are extracted from historical atmospheric pollutant data, and auxiliary data are incorporated into a proposed model to improve the performance, so that the concentration prediction method is realized.

The invention discloses a fundus image optic disc segmentation method and a system for assisting glaucoma screening, and relates to the technical field of image information processing. The fundus image optic disc segmentation method comprises the steps that a plurality of fundus images are collected and preprocessed, and a training image sample set and a verification image sample set are obtained;training of a constructed W-Net-Mcon full convolutional neural network by using the training image sample set to obtain an optimal W-Net-Mcon full convolutional neural networkis carried out; preprocessing the fundus image to be segmented, and inputting the preprocessed fundus image to be segmented into the optimal W-Net-Mcon full convolutional neural network to obtain a prediction target result image; Processing prediction target result graph by utilizing polar coordinate inverse transformation and ellipse fitting to obtain final segmentation result so as to obtain cup-to-disk ratio and finally obtain glaucoma preliminary screening result. According to the method, image semantic information can be effectively extracted in a multi-size mode, fusion of features of different levels, fusion of global features and detail features and encouragement of feature multiplexing are carried out, gradient back propagation is improved, and the image segmentation precision is improved.

The invention provides a small area-level ultra-short-term load prediction and visualization method based on a Long Short-Term Memory (LSTM) network, the method comprising: step 1: an input and outputvariable of the model is determined; step 2: The input and output data sets are separately preprocessed; step 3: the depth LSTM load prediction model is constructed and the random search method is used to find the appropriate hyperparameters until the test set prediction error is minimized. step 4: The t-SNE visualization technology is used to visually characterize the network hidden layer vector, and the correlation coefficient heat map is formed according to the hidden layer vector to perform correlation quantitative analysis, thereby reflecting the network's ability to extract feature datafrom the input data. The method aims to utilize the feature extraction capabilities of the deep learning model and the LSTM temporal correlation learning capabilities to achieve higher prediction accuracy than the machine learning model.

The invention belongs to the technical field of image segmentation, in particular to an image segmentation method, a system and an electronic device based on depth learning. The image segmentation method based on depth learning comprises the following steps: step A: normalizing the original image; B, inputting the normalized image into a ResUNet network model, extracting a feature map containing global semantic information from the input image by the ResUNet network model, upsampling the feature map and stacking the feature map to obtain a final feature map; step c: classifying pixel by pixelthe feature map after upsampling and stacking processing, and outputting the image segmentation result. The present application solves the common problem of gradient disappearance in a convolution neural network, and also enables the network to be more easily trained and converged to a better segmentation result.

The invention relates to a power load forecasting method based on a long-short term neural network, which comprises the steps of S1, performing statistical processing on historical load information, historical temperature information and holiday information of a power grid in a forecast place, dividing the historical load information in every seven days of the power grid in the forecast place intoa group, forecasting the load of the forecast day by inputting 14 pieces of neuron information altogether including the historical load information and holiday information of six days before the forecast day and temperature level information and holiday information of the forecast day so as to serve as training samples, wherein the 14 pieces or neuron information and the training samples constitute a training set; S2, substituting the training set into the long-short term neural network to sequentially perform forward calculation and back propagation training, enabling the error of the training set to be reduced below an error threshold, and obtaining a trained long-short term neural network; and S3, inputting neuron information required by the forecast day into the trained long-short term neural network so as to perform load forecasting on each time point of the forecast day.

The invention provides a lithium battery residual life prediction method based on an LSTM neural network and transfer learning. The lithium battery residual life prediction method comprises the following steps: step 1, carrying out data acquisition and data preprocessing; 2, dividing the data into a training set and a test set in proportion; 3, building a source domain LSTM neural network model, inputting the source domain data training set into a neural network for training, and inputting the data of the test set into the neural network for testing; 4, measuring the data difference between the source domain and the target domain by using the maximum mean value difference to obtain the distribution distance between the source domain and the target domain; and step 5, adjusting the source domain network model according to the maximum mean value difference to obtain a target domain network model, migrating source domain network model parameters, and inputting target domain data into themodel to perform residual life prediction. By migrating the structure and parameters of the network model, the network training time is reduced, and the efficiency is improved.

The invention provides a method for quickly identifying EEG (Electroencephalogram) signals by using a dense deep convolutional neural network, and designs a convolutional neural network suitable for motor imagery EEG signals by combining the characteristics of time and space characteristics of the motor imagery EEG signals and using a characteristic connection method in the convolutional neural network. The convolutional neural network designed by the invention can extract the time and space features at the same time, and the outputs between different convolutional layers are connected with each other, so that the number of weights is reduced, and the purposes of overfitting resistance and feature reuse are achieved. The method comprises the steps of firstly, the filtered and resampled original data is inputted into the dense deep convolutional neural network, then the parameters of each layer of the network are updated through a back propagation and random gradient descent algorithm,finally, the network is tested, the test data is inputted into the trained network, and an output result is analyzed. Compared with a Shallow ConvNet method proposed in 2017, the signal identificationaccuracy and the kappa value are improved by 5% and 0.066%.

The invention relates to a time-space domain correlation prediction method for air pollutant concentration, which comprises the steps of S1, constructing a prediction model based on a residual error network and a convolutional LSTM network by taking PM2.5 as a sample for target pollutant prediction; s2, selecting appropriate training and testing data from the environment monitoring data to complete initialization of the prediction model; s3, training the prediction model stage by stage to obtain a neural network prediction model capable of accurately predicting PM2.5; s4, selecting hyper-parameters (the number of layers, the number of nodes and the learning rate) of the model by utilizing the verification set until the model is optimal; and S5, carrying out urban PM2.5 prediction by utilizing the verified prediction model. Compared with the prior art, the method has the advantages that the convolutional LSTM network is used as a middle layer, deep space-time association feature extraction is performed on spatial features extracted by the bottom ResNet network, accordingly, the prediction performance of the network model can be improved, the hidden state of the convolutional LSTM can be received by the aid of the full connection layer, and a final prediction result can be generated.

The invention discloses a single image fine rain removal method based on a depth convolutional neural network. Firstly, through carrying out background texture structure extraction, nonlinear mapping and rain line area restoration on an inputted rain graph by an initial rain removal network, an initial clear rain-free image is finally obtained; the initial clear rain-free image and the original image are inputted to a fine rain removal network with a single convolutional layer at the same time, more details in the background area are thus restored, and a high-definition rain removal image is finally obtained. Through adopting a caffe framework, the initial rain removal network and the fine rain removal network are trained, parameters of each convolutional layer are obtained precisely, fine rain removal processing is carried out on a rain image, and compared with the traditional convolutional neural network rain removal method, the method of the invention can obtain a higher-quality rain-free image, the practicability is strong, and the method can be widely applied to more scenes.

The invention discloses a text classification method based on a capsule network. The problems that in the prior art, the overall precision is not high, the applicability is not high, a large amount ofimportant information is lost in the feature extraction process, and the relation between the local part and the overall part in a text is ignored are solved. The method comprises the following steps: 1, nodes in a capsule network being capsules consisting of a group of neurons, executing complex internal calculation on input by using matrix capsules, outputting instantiation parameters from results in a matrix form, and meanwhile, outputting an activation value of each capsule is output; 2, calculating between two adjacent layers in the capsule network through an EM routing algorithm; representing a higher-dimensional concept through Gaussian cluster, and each activation capsule selecting a capsule of the next layer as a father node through an iterative routing process, so that link prediction is realized between two adjacent layers of networks; and 3, training a weight parameter of the full connection layer, and calculating the prediction probability of the real label by using a softmax activation function.

The invention discloses a bridge health evaluation method, and relates to the technical field of bridges. The bridge health evaluation method includes acquiring timing-sequence based bridge health monitor data from a bridge health evaluation system, and generating a training sample set on the basis of pre-divided bridge health state levels; performing training by the training sample set to acquirea neural network model; inputting the newly-collected bridge health monitor data into the neural network model, and outputting the predicated bridge health state level. The bridge health evaluation method has the advantages that the newly-collected bridge health monitor data are input into the neural network model, the predicated bridge health state level is outputted, and the bridge health stateis evaluated and predicated effectively.

The invention relates to a hyperspectral remote sensing image classification method based on a double attention mechanism. Compared with the prior art, the hyperspectral remote sensing image classification method solves the defects that hyperspectral remote sensing image training samples are limited and the classification precision is not ideal. The method comprises the following steps: obtaining a training sample; constructing and training a stack auto-encoder; constructing a hybrid dense network; training the hybrid dense network; obtaining a to-be-classified sample and carrying out dimension reduction processing; and obtaining a hyperspectral remote sensing image classification result. According to the method, an ideal hyperspectral remote sensing image classification result can be obtained under the condition that the number of training samples is smaller.

The invention relates to a method for segmenting a liver and a focus thereof in a medical image, and the method comprises the steps: firstly carrying out the screening and integration preprocessing ofabdominal CT image data, dividing the abdominal CT image data into a plurality of data sets with different purposes, building a new neural network, and carrying out the initial training through employing small image data; then, storing the trained model, carrying out secondary training by using the original image and a new data enhancement mode, carrying out expansion and corrosion processing onthe predicted image, and evaluating by using a medical evaluation index; through the model prediction results trained by DL, GDL and TL loss functions, adding and averaging the prediction results of the three loss models to form a fusion feature; finally, modifying the network, wherein the three loss models are fused in a single network for training prediction. End-to-end training test can be carried out, liver and focus can be identified at the same time with high precision and high speed, doctors are effectively helped to identify CT images, time and energy consumed by doctors are greatly reduced, and the probability of misdiagnosis is reduced.

The invention discloses a segmentation method for an inner membrane in a blood vessel of an intravascular unltrasound image. The method comprises the steps of S1, collecting intravascular unltrasoundimage data, and marking a central inner membrane area in each intravascular unltrasound image; S2, performing polar coordinate transformation on a training set image; S3, computing an edge distance map of a mark chart; S4, setting a sliding clamping window and determining the size of the window and a sliding step length; S5, building and training a FusionNet deep learning segmentation model; and S6, inputting new data and acquiring a segmentation result. The method provided by the invention, the inner membrane area in the blood vessel can be quickly, accurately and effectively extracted.

The invention discloses a rapid magnetic resonance imaging method based on an AR2 U-Net neural network. According to the method, an existing R2U-Net convolutional neural network is improved; the method is based on an R2U-Net convolutional neural network. An attention gate module is added, an AG training model is used for implicitly learning and suppressing an unrelated region in an input image, and meanwhile, obvious features useful for a specific task are highlighted, so that the AR2 U-Net convolutional neural network only needs a small amount of training data under the condition of reconstructing an image with the same quality. The problem that the swing amplitude of an optimization loss function is too large in the updating process is solved. According to the method, an Adam optimization algorithm is adopted to replace a conventional SGD optimization algorithm, the convergence rate of the convolutional network can be further increased, the problem that training is ended too early can be effectively prevented, and for learning rate processing, a polynomial attenuation strategy is adopted, so that learning can be stably reduced, and the reduction speed is higher along with increase of the number of turns.

The application discloses a GAN image processing method and system. The method comprises the steps of receiving random noise, and generating a generative image by using a generative network based on an improved LSGAN; receiving a real image, and performing image gradient transformation of different degrees on the real image and the generative image respectively to obtain a transformed image set; and inputting transformed images in the transformed image set respectively into channels of a multi-channel convolutional network of a discrimination network, and extracting and fusing features to obtain an output result. In this application, the random noise is input to the generative network to generate the generative image, then the real image is received, the image gradient transformation of different degrees are respectively performed on the real image and the generative image to obtain the transformed image set, the transformed images in the transformed image set are input respectively into the channels of the multi-channel convolutional network of the discrimination network, extraction and fusion of features are performed to obtain the output result, the network is allowed to have a good generalization ability, the phenomenon of gradient disappearance avoided, and more realistic pictures with higher quality can be output.

The invention discloses a macular location method based on improved Faster R-CNN, which includes collecting training samples, constructing a network model, training the network model, constructing a detection model and performing macular detection and location. The invention utilizes the improved Faster R-CNNto achieve effective localization of macular region, reduces the influence of optic disc and blood vessel on macular region and has strong anti-noise ability, which greatly improves the accurate localization of macular region and lays a foundation for the subsequent analysis and processingof fundus images.

The invention discloses an image multi-subtitle automatic generation method based on a multiscale hierarchical residual network, and adopts an improved funnel network to capture multiscale target information. Firstly, when a funnel framework network is constructed, a densely connected polymerization residual block is put forward, and residual LSTM (Long Short Term Memory) is further put forward inorder to solve the problems of gradient vanishing and gradient explosion. By use of the method, high experiment performance is obtained, and the method has an obvious advantage on multi-subtitle taskacquisition.